By Online With AI · Published · 3 min read
An AI agent can choose steps and tools as it works towards a goal. A workflow follows a more defined route. The commercially useful question is not which sounds more advanced, but where judgement is actually needed.
Start with the decision, not the label
Consider a document received through a form. Extracting a reference number and routing it to an agreed folder can be a fixed workflow. Investigating a question across several approved sources, deciding which evidence is missing and preparing a reasoned response may benefit from an agent.
Anthropic’s engineering guide distinguishes predefined workflows from agents that direct their own processes and tool use. It recommends starting with simple solutions and adding complexity when needed. That distinction is useful even when the underlying model comes from another provider.
A practical selection test
| Task characteristic | A sensible starting point |
|---|---|
| Rules and steps are known | Deterministic workflow |
| One step needs language interpretation | Workflow with a model-assisted step |
| The next step depends on discovered evidence | Bounded agent with approved tools |
| The outcome commits money or affects people materially | Explicit human decision before action |
This is a design aid, not a universal classification. The same process can combine all four. For example, a fixed intake step can feed an agent that drafts an analysis, followed by a person approving a deterministic update.
Give the agent a job description
Specify the goal, permitted sources, tools, stopping conditions and escalation route. “Manage the pipeline” is too broad. “Prepare a weekly account summary from these approved records; flag missing information; do not change deal values” is testable.
Tool permissions should reflect that distinction. Read access to a CRM does not imply write access. Permission to create a draft does not imply permission to send it. Even a highly capable model should operate inside explicit application controls.
Measure the complete attempt
Agents may make several calls before producing an output. A cheap individual call can become an expensive attempt if it loops or repeatedly retrieves the same material. Set limits on time, tool calls and spend, and report the reason when a limit is reached.
Evaluate completed tasks rather than isolated answers. Did the system use current evidence? Did it respect permissions? Did it avoid creating duplicate work? How long did the reviewer spend checking it? These questions reveal whether autonomy improved the process or simply moved the effort elsewhere.
A good first pilot is deliberately bounded
Choose a narrow internal task with recoverable outputs. A briefing assistant is often easier to evaluate than an agent making external commitments. Keep a comparison set of tasks completed by the existing process, and have the same subject-matter owner review both.
Include awkward cases: contradictory records, missing access, a document that contains misleading instructions and a tool that fails. The agent should return useful uncertainty or escalate, rather than improvising a success story.
Autonomy can increase as evidence supports it. It should not increase merely because a new model was released. Read our agent testing guide before expanding a pilot into live operations.